Pay Attention to What and Where? Interpretable Feature Extractor in Vision-based Deep Reinforcement Learning